# Instrumented chat diagnostic: preserved failed artifact-search initial

Run [40635108320](https://github.com/paperclipai/paperclip/actions/runs/40635108320) measured one real human request: **“Find the final website files we delivered last week.”** The original result is **FAILED**, with cleanup passed and evidence manifest reporting no leaks or missing files. One human query and one native run were observed; the native succeeded, while the feature grader failed two outcome assertions.

This is an explicitly instrumented diagnostic variant, source `3d7dfc996dd515549a8a96c4db14c96e0b16676b`, based on production `39e811c9c5226979fce42e052a27cfb28301018a`. Its native binary and build were authenticated separately. Model: `openrouter/deepseek/deepseek-v4.1-flash-1017`, through OpenCode/OpenRouter. It is **not a clean-production pass**, and does not change clean-production coverage: **21 distinct measured, 4 passed, 17 failed, 223 not measured**, out of 244.

## What the diagnostic actually established

| Evidence | Observed outcome |
| --- | --- |
| Actual feature result | FAILED: `exact-accessible-final-and-download` and `history-not-rewritten-by-search`; original status and finite initial allowance preserved. |
| Timing sample | 15 provider-observed completed Paperclip MCP tools, zero failed tools or request timeouts. 253 closed timing records stayed below the 256-record cap. This sample does not establish universal transport health or explain earlier failures. |
| Search answer | Correctly selected the October 21 final source and distinguished the September 30 older final and unapproved draft. A rendered source link and a new current-chat attachment were returned. |
| Historical state | All four original task records retained their compared fields unchanged. The grader nevertheless rejects the new current-chat attachment because its mutation check is broader than historical source entities. |
| Download proof limitation | A successful server preparation receipt binds the returned copy to the expected filename, 115-byte size and original SHA. The observer independently downloaded the original source file, **not the returned copy**. A full verified-copy pass cannot be inferred. |
| Oracle limitation | The grader requires the original attachment URL rather than a verified response copy, and treats the explicitly excluded older file's date as a conflicting selected delivery date. These are future correction candidates, not a retrospective pass. |

The frozen outcome requires usable source/download links, requested time context and restricted-content exclusion. It does not prohibit a truthful verified response copy. A future correction must independently download that actual returned copy, verify exact expected bytes and source identity/provenance, preserve historical task/comment/attachment state, and retain access-denial checks. No case has been regraded or rerun here; this initial remains failed and one final allowance remains subject to explicit later admission.

## Additive interpretation of the preceding budget failure

The earlier `budgets-costs.resolve-budget` final remains FAILED and exhausted. Retained UI evidence shows a real bound budget question with incident context and a $10 choice, despite empty prose. That is a user-facing control, not proof that the agent supplied no response. The fixture used a separate board incident control without answering the pending question, so the actual question/response-wake continuation was not verified. A future fixture must answer the visible bound question and declare every human/native continuation, while retaining board-only policy authority and remaining hold gates. This does not alter the [immutable original first-four report](https://d1p6rlowie26tp.cloudfront.net/runner-e2e/campaigns/gha-40630370749-1-first4-report/).

## What actually works on the current production baseline

| Actual human request | Observed end-to-end state proof | Existing evidence page |
| --- | --- | --- |
| Have Morgan report to Riley from now on. | Saved manager change and organization view; relevant assignment and company invariants retained, 9/9 checks. | [Pilot](https://d1p6rlowie26tp.cloudfront.net/runner-e2e/campaigns/gha-40620786554-1-five-report/) |
| Archive the Launch project but keep its history. | Actual archived project with readable retained history, 10/10 checks. | [Pilot](https://d1p6rlowie26tp.cloudfront.net/runner-e2e/campaigns/gha-40620786554-1-five-report/) |
| Save a short implementation plan on the website task. | Versioned plan saved on the intended task, actual document link and substantive brief facts retained, 9/9 checks. | [Twelve-case report](https://d1p6rlowie26tp.cloudfront.net/runner-e2e/campaigns/gha-40624431756-1-full12-report/) |
| Fix the confusing pricing paragraph and resolve my comment. | Actual anchored document edit and correct annotation resolution, unrelated feedback preserved, 9/9 checks. | [Twelve-case report](https://d1p6rlowie26tp.cloudfront.net/runner-e2e/campaigns/gha-40624431756-1-full12-report/) |

These four are measured passes on production39e. [The full 244-scenario coverage inventory and prerequisite lists](https://d1p6rlowie26tp.cloudfront.net/runner-e2e/campaigns/gha-40584967834-1-final34-report/) remain inspectable in the preceding historical report. Its old outcomes are historical, not extra current-baseline passes. Current measured outcomes and counts are below.

## Coverage and next measurements

The diagnostic's one failed initial is separate from the 21 clean-production scenarios above. Historical production/model results remain historical. The existing 40-case run has resumed bounded waves on its unchanged source after this diagnostic; pending jobs are not measured outcomes. Further actual artifacts will update coverage only after authenticated import. No cost of zero is inferred from unpriced usage.